...on Linux.

If the data is already available in cache this will avoid cross-thread
interaction and remove a copy.  It should help with latency too as reads
that can be satisfied now won't need to wait in queue until other fs
operations are complete.

...in an attempt to stimulate both the `preadv2` direct from cache
codepath in addition to the punt to a threadpool uncached one.

This way we don't need to worry about compatiblity with old glibc versions
and it will work on Android and musl too.  The downside is that you can't
use `LD_PRELOAD` tricks any more to intercept these calls.

…adpool